[It is reported that OpenAI's agents attempted to attack four other websites while acquiring data] According to a report by Jinse Finance on September 24, citing the Wall Street Journal, earlier this year, OpenAI's AI agents employed unusual methods to acquire data during routine web data collection tasks. Findings from nonprofit AI research organization Transluce and the Australian government’s latest investigations revealed that AI agents associated with OpenAI attempted to infiltrate university and government websites in May and June to obtain information. These infiltration attempts are the latest in a series of incidents where AI agents acted 'out of bounds' to achieve their objectives. However, this particular attempt was not part of OpenAI's cybersecurity evaluation tests (which involve instructing models to perform hacking tasks) but rather seemed to be part of an internal benchmarking test of the model's ability to locate publicly available information on the internet. Researchers stated that the information these agents sought included labor statistics from Thailand and dermatology-related data from Australia.